A method and system for assisting statistical analysis of student learning based on a multimodal large model.

Abstract

This invention discloses a method and system for assisting in the statistical analysis of learning outcomes based on a multimodal large model, relating to the field of learning outcome assistance. The method includes: constructing learning outcome evolution sequences for N target objects, selecting K evolution pattern clusters, wherein each cluster contains several learning outcome evolution sequences, determining the evolution label of each evolution pattern cluster, extracting the learning outcome state vectors of the first T time windows, wherein the first T time windows represent the minimum number of time windows that can reliably distinguish the evolution pattern clusters, using the learning outcome state vectors of the first T time windows as feature inputs, and using the evolution label of the evolution pattern cluster to which it belongs as the supervision target, until N supervised training samples are obtained, and providing personalized learning outcome intervention assistance to the object to be identified. This invention only needs to collect the learning outcome state vectors of the shortest time period of the first T time windows to obtain its possible evolution labels, thereby realizing personalized assistance based on learning outcome statistical analysis.

Description

Technical Field

[0001] This invention relates to the field of learning support, specifically a method based on a multimodal large model for assisting in statistical analysis of learning. Background Technology

[0002] In current educational practice, learning analysis is typically based on statistical summarization of students' learning behavior data (such as homework completion, test scores, platform login records, etc.) over a period of time to help teachers understand the learning status of the class or individuals. Common practices include: calculating weekly / monthly average scores, statistically analyzing the accuracy rate of knowledge points, or plotting periodic performance curves.

[0003] However, such methods often rely on complete or long-term historical data, requiring time until the teaching process is halfway through or even nearing its end before effective judgments can be made. This makes it difficult for teachers to identify potential learning risks or developmental trends in students in the early stages of teaching. Some systems attempt to introduce machine learning models to predict learning progress, but their input features typically use a fixed-length time window (e.g., the first four weeks), failing to consider the differences in behavioral evolution among different student groups, thus limiting the ability to make accurate judgments in the early stages.

[0004] Furthermore, existing clustering or classification methods often treat learning situations as isolated event sequences, lacking an expression of the inherent separability of population evolution patterns over time. Consequently, they cannot dynamically determine an observation time window that can both ensure the reliability of pattern differentiation and be as short as possible, thus limiting the timeliness and accuracy of early intervention. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for assisting in statistical analysis of student learning based on a multimodal large model. This method solves the technical problems mentioned in the background by introducing a minimum number of time windows that can be reliably distinguished.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for assisting statistical analysis of student learning based on a multimodal large model includes the following steps:

[0008] S1. Construct a learning progress sequence for N target objects;

[0009] S2. Based on the learning progress evolution sequences of N target objects, select K evolution pattern clusters; where each cluster contains several learning progress evolution sequences;

[0010] S3. Based on several learning situation evolution sequences, determine the evolution labels of each evolution pattern cluster;

[0011] S4. In each learning situation evolution sequence, extract the learning situation status vector of the first T time windows;

[0012] Among them, the first T time windows represent the minimum number of time windows that can reliably distinguish the evolution pattern clusters, and the Tth time window within the first T time windows represents the first time that the average inter-class distance exceeds the separability threshold.

[0013] S5. Take the learning status vector of the first T time windows as the feature input, and the evolution label of the evolution pattern cluster in which it belongs as the supervision target, until N supervised training samples are obtained.

[0014] S6. Based on N supervised training samples, pre-train a supervised classification model for predicting evolution labels;

[0015] S7. Based on a pre-trained supervised classification model, personalized learning intervention assistance is provided for the objects to be identified.

[0016] In some embodiments, a learning progress sequence for N target objects is constructed, including:

[0017] S1-1. Identify N target objects in the sampling population;

[0018] S1-2. Obtain the multimodal learning parameters of N target objects within a predefined time window;

[0019] S1-3. Characterize the multimodal learning parameters and concatenate them into a learning status vector of the target object within the time window;

[0020] S1-4. Obtain the learning status vector of the target object in M ​​consecutive time windows, and assign a window number based on the time window to each learning status vector.

[0021] S1-5. Based on the successively increasing window number, the M learning status vectors are arranged in ascending order to obtain the learning status evolution sequence of N target objects.

[0022] In some embodiments, K evolution pattern clusters are selected based on the learning evolution sequences of N target objects, including:

[0023] S2-1. Define N target objects as N clusters;

[0024] S2-2. Identify the two clusters with the minimum sequence distance among N clusters and merge them into a new cluster;

[0025] S2-3. Calculate the average connection distance between the new cluster and the remaining clusters;

[0026] The average connection distance is defined as the average of the pairwise sequence distances between all target objects in a new cluster and all target objects in another cluster in the learning evolution sequence.

[0027] S2-4. Successively select two clusters with the minimum average connection distance and iteratively merge them into a new cluster until N clusters are merged into a single cluster.

[0028] S2-5. During the iterative merging into a single cluster, construct a hierarchical clustering tree that represents the hierarchical relationship of the learning situation evolution sequence clustering.

[0029] S2-6. Anchor K of the aforementioned evolution pattern clusters on the hierarchical clustering tree.

[0030] In some embodiments, identifying two clusters with the minimum sequence distance among N clusters and merging them into a new cluster includes:

[0031] S2-2-1, Anchoring the learning situation evolution sequence contained in each of the N clusters;

[0032] S2-2-2. In each cluster, calculate the sequence distance between any two learning situation evolution sequences to obtain G sequence distances; where G = N × (N-1) / 2;

[0033] S2-2-3. Select the minimum sequence distance from the G sequence distances;

[0034] S2-2-4. Based on the minimum sequence distance, anchor the two clusters with the minimum sequence distance among N clusters;

[0035] S2-2-5. Merge the two clusters with the minimum sequence distance into the new cluster.

[0036] In some embodiments, during the iterative merging into the new cluster, a hierarchical clustering tree representing the hierarchical relationship of the learning situation evolution sequence is constructed, including:

[0037] S2-5-1. During each iteration of merging, the merging operation is assigned a merging stage number that increases sequentially according to the order in which the merging occurs; the value range of the merging stage number is 1 to N-1.

[0038] S2-5-2. Based on the merging stage number, the new cluster generated by this merging is marked as an internal node of the hierarchical clustering tree, and its corresponding average connection distance is marked as the node height of this internal node.

[0039] S2-5-3. Arrange all internal nodes from bottom to top based on the merging stage number to form the hierarchical clustering tree.

[0040] In some embodiments, anchoring K clusters of evolution patterns on a hierarchical clustering tree includes:

[0041] S2-6-1. Extract the minimum sequence distance or minimum average connection distance of each internal node;

[0042] S2-6-2. Starting from the minimum sequence distance, generate a merge distance sequence by arranging the minimum average connection distances in ascending order based on the merge stage number;

[0043] S2-6-3. Calculate the distance increment between adjacent distances in the merged distance sequence, and each distance increment inherits the merging stage number of the previous distance;

[0044] S2-6-4. Based on the inherited merging stage number, arrange the distance increments in ascending order to generate a distance increment sequence;

[0045] S2-6-5. Traverse the distance increment sequence in ascending order of the merging stage number and compare each distance increment with a set threshold.

[0046] If there is an H-th distance increment that is greater than a set threshold, then anchor the H-th merging stage number corresponding to this distance increment on the hierarchical clustering tree;

[0047] S2-6-6, Anchor the K clusters corresponding to the Hth merging stage number, and define the K clusters as K evolutionary pattern clusters;

[0048] Wherein, the K evolutionary pattern clusters represent all the clusters retained after the Hth merging operation is completed, and the number of all clusters is K=NH.

[0049] In some embodiments, the evolution labels of each evolution pattern cluster are determined based on several learning evolution sequences, including:

[0050] S3-1. Select several learning situation evolution sequences from any evolution pattern cluster;

[0051] S3-2. For each learning situation evolution sequence, extract the window number of the corresponding M time windows;

[0052] S3-3. Using the window number as an index, query several learning status vectors located at the same window number in each learning status evolution sequence;

[0053] S3-4. Pool several learning status vectors located in the same window number to obtain the pooled status vector of this window number.

[0054] S3-5. Repeatedly pool the learning status vectors of M window numbers until M pooled status vectors are obtained.

[0055] S3-6. Calculate the L2 norm of each of the M pooling state vectors;

[0056] S3-7. Arrange the M L2 norms in ascending order according to the window number to generate the learning intensity sequence of each evolution pattern cluster;

[0057] S3-8. Calculate the trend slope, mean of the first segment, mean of the second segment, and variance of each learning intensity sequence.

[0058] S3-9. Determine the evolution label for each evolution pattern cluster according to the predefined mapping rules of the trend slope, the mean of the first segment, the mean of the second segment, and the variance of the sequence.

[0059] In some embodiments, the learning status vectors for the first T time windows are extracted from each learning progress evolution sequence, including:

[0060] S4-1. Obtain the M pooling state vectors for each of the K evolutionary pattern clusters;

[0061] S4-2, anchoring M time windows corresponding to M pooling state vectors;

[0062] S4-3. Align the M pooling state vectors of each of the K evolutionary mode clusters according to the window number of the time window;

[0063] S4-4. Select the time window with the same window number as the target window;

[0064] S4-5. Calculate the inter-class distance between the pooling state vectors of any two clusters when they are in the target window, until D inter-class distances are generated; where D = K × (K-1) / 2;

[0065] S4-6. Obtain the average inter-class distance of the D inter-class distances and compare the average inter-class distance with the separability threshold;

[0066] S4-7. Traverse M time windows from 1 to M, and determine the window number of the time window in which the average inter-class distance first exceeds the separability threshold as T.

[0067] S4-8. In N sequences of learning progress evolution, extract the learning status vectors for the first T time windows.

[0068] In some of these embodiments, based on a pre-trained supervised classification model, personalized learning intervention assistance is provided for the object to be identified, including:

[0069] S7-1. Obtain the learning status vector of the object to be identified in the first T time windows;

[0070] S7-2. Input the learning status vector of the first T time windows into the pre-trained supervised classification model and output the learning prediction label of the object to be identified.

[0071] S7-3. Based on the learning progress prediction labels of the target object, provide personalized learning progress intervention and assistance to the target object.

[0072] This invention provides a method for assisting in statistical analysis of student learning based on a multimodal large model, which has the following beneficial effects:

[0073] This invention constructs a learning evolution sequence based on students' multimodal learning behaviors (such as login frequency, answer accuracy, number of interactions, etc.) within a continuous time window; through sequence clustering, students with similar learning evolution trajectories are grouped into the same evolution pattern cluster, and based on the activity change trend of each cluster throughout the semester, they are assigned evolution labels such as stable, continuously improving, and fluctuating.

[0074] Furthermore, the average inter-class distance between the group learning performance (i.e., pooled state vector) of each evolution pattern cluster is calculated under each time window. Starting from the first window, the detection is performed window by window. The time window number that first makes the distance exceed the preset separability threshold is determined as T, thereby ensuring that the first T time windows are the earliest time points that can reliably distinguish different evolution patterns. The early learning parameters extracted in this way and the corresponding evolution labels form training samples for training the classification model.

[0075] Ultimately, when a new student enters the system, it is only necessary to collect their learning status vectors from the previous T time windows. The trained model can then predict their evolution label. Based on this, we can understand the student's potential learning trajectory, adjust teaching arrangements in a timely manner, and send relevant learning information to the parents' terminal, so that families can keep track of the student's learning progress at school.

[0076] Secondly, this invention provides a system based on a multimodal large model for assisting in statistical analysis of student learning, comprising:

[0077] The sequence construction module is used to construct the learning progress evolution sequence of N target objects;

[0078] The cluster selection module is used to select K evolution pattern clusters based on the learning evolution sequences of N target objects; where each cluster contains several learning evolution sequences.

[0079] The label determination module is used to determine the evolution labels of each evolution pattern cluster based on several learning evolution sequences;

[0080] The vector extraction module is used to extract the learning status vectors of the first T time windows in each learning progress evolution sequence;

[0081] The sample construction module is used to take the learning status vector of the first T time windows as the feature input and the evolution label of the evolution pattern cluster to which it belongs as the supervision target, until N supervised training samples are obtained.

[0082] The pre-training module is used to pre-train a supervised classification model for predicting evolving labels based on N supervised training samples.

[0083] The evolution classification module is used to provide personalized learning intervention assistance for the objects to be identified based on the pre-trained supervised classification model.

[0084] Compared with the prior art, the beneficial effects of the system based on a multimodal large model for assisting statistical analysis of learning conditions in this invention are the same as the beneficial effects of the method based on a multimodal large model for assisting statistical analysis of learning conditions described above, so they will not be repeated here. Attached Figure Description

[0085] Figure 1 This is a flowchart illustrating a method for assisting statistical analysis of student learning based on a multimodal large model according to the present invention.

[0086] Figure 2 This is a schematic diagram of an embodiment of the evolutionary pattern cluster described in this invention;

[0087] Figure 3 This is a schematic diagram of another embodiment of the evolutionary pattern cluster described in this invention;

[0088] Figure 4 This is a schematic diagram illustrating the process of extracting the learning status vector for the first T time windows as described in this invention.

[0089] Figure 5 This is a structural block diagram of a system based on a multimodal large model for assisting in statistical analysis of student learning, according to the present invention. Detailed Implementation

[0090] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] Example 1: Please refer to Figures 1 to 4 This invention provides a method for assisting statistical analysis of student learning based on a multimodal large model, comprising the following steps:

[0092] S1. Construct a learning progress sequence for N target objects;

[0093] S2. Based on the learning progress evolution sequences of N target objects, select K evolution pattern clusters; where each cluster contains several learning progress evolution sequences;

[0094] S3. Based on several learning situation evolution sequences, determine the evolution labels of each evolution pattern cluster;

[0095] S4. In each learning situation evolution sequence, extract the learning situation status vector of the first T time windows;

[0096] Among them, the first T time windows represent the minimum number of time windows that can reliably distinguish the evolution pattern clusters, and the Tth time window within the first T time windows represents the first time that the average inter-class distance exceeds the separability threshold; that is, the earliest time window can be determined through the Tth time window, which is the shortest time period from the first time window to the earliest time window.

[0097] S5. Take the learning status vector of the first T time windows as the feature input, and the evolution label of the evolution pattern cluster in which it belongs as the supervision target, until N supervised training samples are obtained.

[0098] S6. Based on N supervised training samples, pre-train a supervised classification model for predicting evolution labels;

[0099] Specifically, the supervised model can be a deep neural network model suitable for multimodal temporal classification tasks, such as an attention-based Transformer model or an LSTM network.

[0100] For example, the supervised model can be a Transformer or LSTM model; it takes the learning status vector of the first T time windows as input, outputs the predicted probability of each evolution label, and drives parameter optimization through cross-entropy loss, thereby learning the mapping relationship between the learning status vector of the first T time windows and the evolution label.

[0101] S7. Based on a pre-trained supervised classification model, personalized learning intervention assistance is provided for the objects to be identified.

[0102] In this embodiment, by constructing a multimodal learning evolution sequence of the target object, clustering them into an evolution pattern cluster with consistent behavioral evolution, and generating evolution labels based on the group trajectory; further, by utilizing the separability between clusters to determine the shortest time window T for reliable differentiation, supervised samples are constructed from the learning state vectors of the previous T time windows and the corresponding evolution labels, and a time-series classification model is trained, thereby enabling the discrimination of the long-term evolution trend of new objects based on early learning data, thus allowing for personalized learning intervention assistance.

[0103] Specifically, in this embodiment, S1 includes:

[0104] S1-1. Identify N target objects in the sampling population;

[0105] S1-2. Obtain the multimodal learning parameters of N target objects within a predefined time window;

[0106] Specifically, a predefined time window represents the sampling period used to collect and classify learning behavior parameters. Its length can be a fixed duration (such as one week) or divided according to the teaching cycle (such as the learning cycle of a chapter).

[0107] Multimodal learning parameters represent quantitative indicators reflecting learner behavior parameters collected from multiple data sources, including login frequency to online learning platforms, completion rate of course videos, accuracy rate of quiz answers, number of interactions in learning communities, timeliness of assignment submissions, etc.

[0108] S1-3. Characterize the multimodal learning parameters and concatenate them into a learning status vector of the target object within the time window;

[0109] S1-4. Obtain the learning status vector of the target object in M ​​consecutive time windows, and assign a window number based on the time window to each learning status vector.

[0110] Specifically, the window number represents the relative position index of the time window on the time axis, which can be determined by the start or end timestamp of the time window on the time axis; specifically, the earliest time window is assigned the number 1, and each subsequent time window is incremented by 1 in chronological order, forming a continuous integer sequence from 1 to M.

[0111] S1-5. Based on the successively increasing window number, the M learning status vectors are arranged in ascending order to obtain the learning status evolution sequence of N target objects.

[0112] In this embodiment, by characterizing and concatenating the multimodal learning parameters collected for each target object within M consecutive predefined time windows into learning status vectors for the corresponding windows, and assigning incremental window numbers according to the chronological order, a learning evolution sequence with time series as its organizational structure is formed, so that the original scattered behavior logs are transformed into a structured vector sequence with a clear time series index.

[0113] Specifically, in this embodiment, S2 includes:

[0114] S2-1. Define N target objects as N clusters;

[0115] S2-2. Identify the two clusters with the minimum sequence distance among N clusters and merge them into a new cluster;

[0116] S2-3. Calculate the average connection distance between the new cluster and the remaining clusters;

[0117] The average connection distance is defined as the average of the pairwise sequence distances between all target objects in a new cluster and all target objects in another cluster in the learning evolution sequence.

[0118] For example, suppose there are currently two clusters. and ,in Includes target objects A and B. Includes target objects C and D; assuming the pairwise sequence distances between the learning progress evolution sequences are: d(A,C)=2.0, d(A,D)=1.6, d(B,C)=1.8, d(B,D)=2.2, then the cluster... and The average connection distance between them is the average of all inter-cluster distances:

[0119] AD= ;

[0120] Where AD represents the average connection distance.

[0121] S2-4. Successively select two clusters with the minimum average connection distance and iteratively merge them into a new cluster until N clusters are merged into a single cluster.

[0122] S2-5. During the iterative merging into a single cluster, construct a hierarchical clustering tree that represents the hierarchical relationship of the learning situation evolution sequence clustering.

[0123] S2-6. Anchor K of the aforementioned evolution pattern clusters on the hierarchical clustering tree.

[0124] In this embodiment, by calculating the sequence distance between the learning situation evolution sequences, and successively merging the two clusters with the smallest distance based on the average distance of cross-cluster sequence pairs, the average distance and order of each merge are recorded during the merging process, thereby constructing a clustering hierarchical structure representing the learning situation evolution sequences of N target objects, thus providing a structured representation for selecting K evolution pattern clusters.

[0125] Furthermore, step S2-2 also includes:

[0126] S2-2-1, Anchoring the learning situation evolution sequence contained in each of the N clusters;

[0127] The learning situation evolution sequence represents the trajectory of the target object's multimodal learning situation status changing over time within a continuous time window;

[0128] S2-2-2. In each cluster, calculate the sequence distance between any two learning situation evolution sequences to obtain G sequence distances; where G = N × (N-1) / 2;

[0129] The formula for calculating the sequence distance is:

[0130] ;

[0131] in, and Let M represent the learning progress sequence of the i-th and j-th target objects, respectively, and M represent the total number of time windows. and Let represent the learning status vector of the two learning progress sequences at the t-th time window. This indicates the calculation of the Euclidean distance between the state vectors of two students.

[0132] S2-2-3. Select the minimum sequence distance from the G sequence distances;

[0133] S2-2-4. Based on the minimum sequence distance, anchor the two clusters with the minimum sequence distance among N clusters;

[0134] S2-2-5. Merge the two clusters with the minimum sequence distance into the new cluster.

[0135] In this embodiment, the sequence distance is calculated by using the complete learning evolution sequence as the basic comparison unit, and the two closest clusters are identified and merged accordingly, so that the clustering process can capture the evolutionary similarity of learners' multimodal behavior over a continuous time window.

[0136] Furthermore, step S2-5 also includes:

[0137] S2-5-1. During each iteration of merging, the merging operation is assigned a merging stage number that increases sequentially according to the order in which the merging occurs; the value range of the merging stage number is 1 to N-1.

[0138] S2-5-2. Based on the merging stage number, the new cluster generated by this merging is marked as an internal node of the hierarchical clustering tree, and its corresponding average connection distance is marked as the node height of this internal node.

[0139] S2-5-3. Arrange all internal nodes from bottom to top based on the merging stage number to form the hierarchical clustering tree.

[0140] In this embodiment, by assigning a unique merging stage number to each merging operation in the order of occurrence, and using the stage number as a time sequence index, the new cluster and its corresponding average connection distance are encoded as the internal nodes and node heights of the tree, respectively, thereby constructing a hierarchical clustering tree that can completely trace back the clustering merging path and the evolution of inter-cluster distances.

[0141] Furthermore, steps S2-6 also include:

[0142] S2-6-1. Extract the minimum sequence distance or minimum average connection distance of each internal node;

[0143] S2-6-2. Starting from the minimum sequence distance, generate a merge distance sequence by arranging the minimum average connection distances in ascending order based on the merge stage number;

[0144] S2-6-3. Calculate the distance increment between adjacent distances in the merged distance sequence, and each distance increment inherits the merging stage number of the previous distance;

[0145] S2-6-4. Based on the inherited merging stage number, arrange the distance increments in ascending order to generate a distance increment sequence;

[0146] S2-6-5. Traverse the distance increment sequence in ascending order of the merging stage number and compare each distance increment with a set threshold.

[0147] If there is an H-th distance increment that is greater than a set threshold, then anchor the H-th merging stage number corresponding to this distance increment on the hierarchical clustering tree;

[0148] S2-6-6, Anchor the K clusters corresponding to the Hth merging stage number, and define the K clusters as K evolutionary pattern clusters;

[0149] Wherein, the K evolutionary pattern clusters represent all the clusters retained after the Hth merging operation is completed, and the number of all clusters is K=NH.

[0150] In this embodiment, by calculating the distance increment of adjacent items in the merged distance sequence and locating the merging stage number H corresponding to the first significantly increased distance increment according to a preset threshold, the N-H clusters remaining after the Hth merging of the hierarchical clustering tree are selected as K evolution pattern clusters, so that the learning situation evolution sequence in each cluster has high similarity in the overall behavioral trajectory, and there are quantifiable distinguishing boundaries between different clusters.

[0151] Specifically, in this embodiment, step S3 includes:

[0152] S3-1. Select several learning situation evolution sequences from any evolution pattern cluster;

[0153] S3-2. For each learning situation evolution sequence, extract the window number of the corresponding M time windows;

[0154] S3-3. Using the window number as an index, query several learning status vectors located at the same window number in each learning status evolution sequence;

[0155] Specifically, the learning status vector is a structured vector formed by normalizing and concatenating the multimodal learning parameters of the target object within a single time window. Each dimension corresponds to quantifiable learning behavior indicators such as login frequency, video viewing completion rate, test answer accuracy, forum interaction frequency, and homework submission timeliness.

[0156] In a specific application scenario, individualized learning behavior evolution curves can be generated based on each student's learning progress sequence and the evolution pattern cluster to which they belong. Combined with the current teaching progress, these curves can be mapped to specific knowledge points, thereby visually presenting each student's mastery of each knowledge point and pushing the data to the associated parent terminal to achieve collaboration between school analysis and home analysis.

[0157] S3-4. Pool several learning status vectors located in the same window number to obtain the pooled status vector of this window number.

[0158] It should be noted that the pooling here is not used to calculate the center vector of the evolution pattern cluster, but rather to treat the learning evolution sequence of multiple target objects within the same evolution pattern cluster as a continuously evolving collective learning behavior trajectory on the time axis. By aggregating several learning state vectors of the cluster within the same time window, the overall learning performance of the group at that time point is obtained. Therefore, the pooled state vector in this embodiment can be understood as the group learning performance representation vector of the evolution pattern cluster within a specific time window.

[0159] S3-5. Repeatedly pool the learning status vectors of M window numbers until M pooled status vectors are obtained.

[0160] S3-6. Calculate the L2 norm of each of the M pooling state vectors;

[0161] Specifically, the L2 norm represents the overall activity level of multimodal learning behavior of the evolutionary pattern cluster within the corresponding time window; the larger the value, the higher the overall learning engagement of learners in the cluster during that period.

[0162] S3-7. Arrange the M L2 norms in ascending order according to the window number to generate the learning intensity sequence of each evolution pattern cluster;

[0163] Specifically, the learning intensity sequence is obtained by arranging the L2 norms of M pooled state vectors in ascending order according to the window number of the time window, thereby representing the changing trend of the overall learning activity level of the corresponding evolution pattern cluster within a continuous time window.

[0164] S3-8. Calculate the trend slope, mean of the first segment, mean of the second segment, and variance of each learning intensity sequence.

[0165] Specifically, the trend slope represents the linear trend of the learning intensity sequence over time, and its regression coefficient is obtained by linearly fitting the sequence using the least squares method.

[0166] Both the mean of the first segment and the mean of the second segment represent the average activity level of the sequence at different stages. The first third of the sequence window can be selected as the first segment and the second third of the sequence window as the second segment.

[0167] Sequence variance represents the degree of fluctuation in the learning intensity sequence, reflecting the stability of the learning behavior of this evolutionary pattern cluster.

[0168] S3-9. Determine the evolution label for each evolution pattern cluster according to the predefined mapping rules of the trend slope, the mean of the first segment, the mean of the second segment, and the variance of the sequence.

[0169] Specifically, in this embodiment, the mapping rule means that: based on the statistical characteristics of the trend slope, the mean of the first segment, the mean of the second segment, and the variance of the sequence of learning intensity, classification is performed by using a preset discrimination threshold; wherein, the discrimination threshold can be determined based on the statistical distribution of historical learning parameters, for example, the absolute value threshold of the trend slope can be set to 0.1, and the sequence variance threshold can be set to 0.8.

[0170] For example:

[0171] If the absolute value of the trend slope is less than 0.1 and the series variance is less than 0.8, then the evolution label is stable.

[0172] If the trend slope is greater than 0.1 and the mean of the later segment is greater than the mean of the earlier segment, then the evolution label is continuously improving.

[0173] If the trend slope is less than -0.1 and the mean of the first segment is greater than the mean of the second segment, then the evolution label is a continuous decline type.

[0174] If the sequence variance is greater than 0.8, the evolution label is fluctuation type.

[0175] In this embodiment, the learning status vectors of multiple target objects within the same evolution pattern cluster are aggregated into a group representation according to a time window, and a learning intensity sequence is constructed based on its L2 norm. Then, the statistical features such as the trend slope, mean and variance of the sequence are extracted, and the sequence is mapped to evolution labels such as stable, continuously improving, continuously declining or fluctuating according to a preset threshold rule, thereby realizing the expression of the evolution trend of the group's learning status.

[0176] Specifically, in this embodiment, step S4 includes:

[0177] S4-1. Obtain the M pooling state vectors for each of the K evolutionary pattern clusters;

[0178] S4-2, anchoring M time windows corresponding to M pooling state vectors;

[0179] S4-3. Align the M pooling state vectors of each of the K evolutionary mode clusters according to the window number of the time window;

[0180] S4-4. Select the time window with the same window number as the target window;

[0181] S4-5. Calculate the inter-class distance between the pooling state vectors of any two clusters when they are in the target window, until D inter-class distances are generated; where D = K × (K-1) / 2;

[0182] In this embodiment, the inter-class distance is preferably the Euclidean distance between pooling state vectors; its calculation formula is:

[0183] ;

[0184] in, and These represent the pooling state vectors of the i-th and j-th evolutionary pattern clusters under the target window, respectively. This represents the dimension of the pooling state vector. and These represent the component values ​​of the two pooled state vectors in the k-th dimension.

[0185] S4-6. Obtain the average inter-class distance of the D inter-class distances and compare the average inter-class distance with the separability threshold;

[0186] S4-7. Traverse M time windows from 1 to M, and determine the window number of the time window in which the average inter-class distance first exceeds the separability threshold as T.

[0187] S4-8. In N sequences of learning progress evolution, extract the learning status vectors for the first T time windows.

[0188] In this embodiment, the average inter-class distance between the group representations of each evolutionary pattern cluster is calculated successively along the time window to quantify the distinguishability of different evolutionary pattern clusters at each time point. Since the pooled state vector represents the group learning behavior of each cluster under the corresponding time window, the larger the average inter-class distance between them, the more significant the overall behavioral differences of different evolutionary patterns at that time, that is, the higher the separability between patterns. Therefore, the time point at which the preset separability threshold is first reached is determined as T, so that the length of the intercepted time window can adaptively reflect the minimum time requirement for different learning evolutionary patterns to be reliably distinguishable in the early stage.

[0189] Specifically, in this embodiment, step S7 includes:

[0190] S7-1. Obtain the learning status vector of the object to be identified in the first T time windows;

[0191] S7-2. Input the learning status vector of the first T time windows into the pre-trained supervised classification model and output the learning prediction label of the object to be identified.

[0192] S7-3. Based on the learning progress prediction labels of the target object, provide personalized learning progress intervention and assistance to the target object.

[0193] It should be noted that personalized learning intervention assistance refers to learning support strategies generated based on learning prediction labels and tailored to the target audience; it is implemented through a pre-defined rule mapping mechanism, specifically including:

[0194] When the predicted label is continuously declining, the process of sending learning plan adjustment suggestions and psychological counseling resources is triggered.

[0195] When the prediction label is "fluctuation type", initiate the task of generating a learning behavior pattern analysis report and recommend time management tools;

[0196] When the predicted label is stable, the intelligent recommendation module for advanced learning content or academic competition information is activated.

[0197] The learning support strategies described above can be driven by a rule engine to enable rule-based intervention decisions based on the students' learning progress.

[0198] Example 2: This Example 2 differs from Example 1 in that it also provides a system based on a multimodal large model for assisting in statistical analysis of student learning. This system is used to implement the above-described method embodiments, and details already described will not be repeated. The terms "module," "unit," and "subunit" used below refer to combinations of software and / or hardware that achieve a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0199] Figure 5 This is a structural block diagram of a system based on a multimodal large model for assisting in statistical analysis of student learning, according to the present invention. The system includes:

[0200] The sequence construction module is used to construct the learning progress evolution sequence of N target objects;

[0201] The cluster selection module is used to select K evolution pattern clusters based on the learning evolution sequences of N target objects; where each cluster contains several learning evolution sequences.

[0202] The label determination module is used to determine the evolution labels of each evolution pattern cluster based on several learning evolution sequences;

[0203] The vector extraction module is used to extract the learning status vectors of the first T time windows in each learning progress evolution sequence;

[0204] The sample construction module is used to take the learning status vector of the first T time windows as the feature input and the evolution label of the evolution pattern cluster to which it belongs as the supervision target, until N supervised training samples are obtained.

[0205] The pre-training module is used to pre-train a supervised classification model for predicting evolving labels based on N supervised training samples.

[0206] The evolution classification module is used to provide personalized learning intervention assistance for the objects to be identified based on the pre-trained supervised classification model.

[0207] In the above system, a sequence construction module is used to construct a learning situation evolution sequence, a cluster selection module is used to select K evolution pattern clusters, each of which contains several learning situation evolution sequences, a label determination module is used to determine the evolution label of each evolution pattern cluster, a vector extraction module is used to extract the learning situation state vector of the first T time windows, a sample construction module is used to obtain N supervised training samples, a pre-training module is used to pre-train the supervised classification model, and an evolution classification module is used to provide personalized learning situation intervention assistance to the object to be identified, thus solving the problem of timeliness and accuracy in early intervention.

[0208] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means.

[0209] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for assisting in statistical analysis of student learning based on a multimodal large model, characterized in that, include: S1. Based on the multimodal learning behaviors of N target objects within a continuous time window, construct a learning evolution sequence for the N target objects, wherein the multimodal learning behaviors include login frequency, answer accuracy rate, and number of interactions; S2. Based on the learning progress evolution sequences of N target objects, select K evolution pattern clusters; where each cluster contains several learning progress evolution sequences; S3. Based on several learning situation evolution sequences, determine the evolution labels of each evolution pattern cluster; S4. In each learning situation evolution sequence, extract the learning situation status vector of the first T time windows; Among them, the first T time windows represent the minimum number of time windows that can reliably distinguish the evolution pattern clusters, and the Tth time window within the first T time windows represents the first time that the average inter-class distance exceeds the separability threshold. S5. Take the learning status vector of the first T time windows as the feature input, and the evolution label of the evolution pattern cluster in which it belongs as the supervision target, until N supervised training samples are obtained. S6. Based on N supervised training samples, pre-train a supervised classification model for predicting evolution labels; S7. Based on a pre-trained supervised classification model, personalized learning intervention assistance is provided for the objects to be identified.

2. The method for assisting statistical analysis of student learning based on a multimodal large model according to claim 1, characterized in that, Construct a learning progress sequence for N target objects, including: S1-1. Identify N target objects in the sampling population; S1-2. Obtain the multimodal learning parameters of N target objects within a predefined time window; S1-3. Characterize the multimodal learning parameters and concatenate them into a learning status vector of the target object within the time window; S1-4. Obtain the learning status vector of the target object in M ​​consecutive time windows, and assign a window number based on the time window to each learning status vector. S1-5. Based on the successively increasing window number, the M learning status vectors are arranged in ascending order to obtain the learning status evolution sequence of N target objects.

3. The method for assisting statistical analysis of student learning based on a multimodal large model according to claim 1, characterized in that, Based on the learning progress evolution sequence of N target subjects, K evolution pattern clusters are selected, including: S2-1. Define N target objects as N clusters; S2-2. Identify the two clusters with the minimum sequence distance among N clusters and merge them into a new cluster; S2-3. Calculate the average connection distance between the new cluster and the remaining clusters; The average connection distance is defined as the average of the pairwise sequence distances between all target objects in a new cluster and all target objects in another cluster in the learning evolution sequence. S2-4. Successively select two clusters with the minimum average connection distance and iteratively merge them into a new cluster until N clusters are merged into a single cluster. S2-5. During the iterative merging into a single cluster, construct a hierarchical clustering tree that represents the hierarchical relationship of the learning situation evolution sequence clustering. S2-6. Anchor K of the aforementioned evolution pattern clusters on the hierarchical clustering tree.

4. The method for assisting statistical analysis of learning based on a multimodal large model according to claim 3, characterized in that, Identify the two clusters with the minimum sequence distance among N clusters and merge them into a new cluster, including: S2-2-1, Anchoring the learning situation evolution sequence contained in each of the N clusters; S2-2-2. In each cluster, calculate the sequence distance between any two learning situation evolution sequences to obtain G sequence distances; where G = N × (N-1) / 2; S2-2-3. Select the minimum sequence distance from the G sequence distances; S2-2-4. Based on the minimum sequence distance, anchor the two clusters with the minimum sequence distance among N clusters; S2-2-5. Merge the two clusters with the minimum sequence distance into the new cluster.

5. The method for assisting statistical analysis of learning based on a multimodal large model according to claim 4, characterized in that, During the iterative merging into the new cluster, a hierarchical clustering tree representing the hierarchical relationship of the learning situation evolution sequence is constructed, including: S2-5-1. During each iteration of merging, the merging operation is assigned a merging stage number that increases sequentially according to the order in which the merging occurs; the value range of the merging stage number is 1 to N-1. S2-5-2. Based on the merging stage number, the new cluster generated by this merging is marked as an internal node of the hierarchical clustering tree, and its corresponding average connection distance is marked as the node height of this internal node. S2-5-3. Arrange all internal nodes from bottom to top based on the merging stage number to form the hierarchical clustering tree.

6. The method for assisting statistical analysis of learning based on a multimodal large model according to claim 5, characterized in that, Anchoring K clusters of the evolution patterns on the hierarchical clustering tree, including: S2-6-1. Extract the minimum sequence distance or minimum average connection distance of each internal node; S2-6-2. Starting from the minimum sequence distance, generate a merge distance sequence by arranging the minimum average connection distances in ascending order based on the merge stage number; S2-6-3. Calculate the distance increment between adjacent distances in the merged distance sequence, and each distance increment inherits the merging stage number of the previous distance; S2-6-4. Based on the inherited merging stage number, arrange the distance increments in ascending order to generate a distance increment sequence; S2-6-5. Traverse the distance increment sequence in ascending order of the merging stage number and compare each distance increment with a set threshold. If there is an H-th distance increment that is greater than a set threshold, then anchor the H-th merging stage number corresponding to this distance increment on the hierarchical clustering tree; S2-6-6, Anchor the K clusters corresponding to the Hth merging stage number, and define the K clusters as K evolutionary pattern clusters; Wherein, the K evolutionary pattern clusters represent all the clusters retained after the Hth merging operation is completed, and the number of all clusters is K=NH.

7. The method for assisting statistical analysis of student learning based on a multimodal large model according to claim 2, characterized in that, Based on several learning evolution sequences, the evolution labels for each evolution pattern cluster are determined, including: S3-1. Select several learning situation evolution sequences from any evolution pattern cluster; S3-2. For each learning situation evolution sequence, extract the window number of the corresponding M time windows; S3-3. Using the window number as an index, query several learning status vectors located at the same window number in each learning status evolution sequence; S3-4. Pool several learning status vectors located in the same window number to obtain the pooled status vector of this window number. S3-5. Repeatedly pool the learning status vectors of M window numbers until M pooled status vectors are obtained. S3-6. Calculate the L2 norm of each of the M pooling state vectors; S3-7. Arrange the M L2 norms in ascending order according to the window number to generate the learning intensity sequence of each evolution pattern cluster; S3-8. Calculate the trend slope, mean of the first segment, mean of the second segment, and variance of each learning intensity sequence. S3-9. Determine the evolution label for each evolution pattern cluster according to the predefined mapping rules of the trend slope, the mean of the first segment, the mean of the second segment, and the variance of the sequence.

8. The method for assisting statistical analysis of student learning based on a multimodal large model according to claim 7, characterized in that, In each learning progress evolution sequence, the learning progress status vectors for the first T time windows are extracted, including: S4-1. Obtain the M pooling state vectors for each of the K evolutionary pattern clusters; S4-2, anchoring M time windows corresponding to M pooling state vectors; S4-3. Align the M pooling state vectors of each of the K evolutionary mode clusters according to the window number of the time window; S4-4. Select the time window with the same window number as the target window; S4-5. Calculate the inter-class distance between the pooling state vectors of any two clusters when they are in the target window, until D inter-class distances are generated; D=K×(K-1) / 2; S4-6. Obtain the average inter-class distance of the D inter-class distances and compare the average inter-class distance with the separability threshold; S4-7. Traverse M time windows from 1 to M, and determine the window number of the time window in which the first average inter-class distance exceeds the separability threshold as T. S4-8. In N sequences of learning progress evolution, extract the learning status vectors for the first T time windows.

9. A method for assisting statistical analysis of student learning based on a multimodal large model according to claim 8, characterized in that, Based on a pre-trained supervised classification model, personalized learning intervention assistance is provided for the objects to be identified, including: S7-1. Obtain the learning status vector of the object to be identified in the first T time windows; S7-2. Input the learning status vector of the first T time windows into the pre-trained supervised classification model and output the learning prediction label of the object to be identified. S7-3. Based on the learning progress prediction labels of the target object, provide personalized learning progress intervention and assistance to the target object.

10. A system based on a multimodal large model for assisting in statistical analysis of student learning, characterized in that, include: The sequence construction module is used to construct a learning evolution sequence of N target objects based on their multimodal learning behaviors within a continuous time window. The multimodal learning behaviors include login frequency, answer accuracy, and number of interactions. The cluster selection module is used to select K evolution pattern clusters based on the learning evolution sequences of N target objects; where each cluster contains several learning evolution sequences. The label determination module is used to determine the evolution labels of each evolution pattern cluster based on several learning evolution sequences; The vector extraction module is used to extract the learning status vectors of the first T time windows in each learning progress evolution sequence; The sample construction module is used to take the learning status vector of the first T time windows as the feature input and the evolution label of the evolution pattern cluster to which it belongs as the supervision target, until N supervised training samples are obtained. The pre-training module is used to pre-train a supervised classification model for predicting evolving labels based on N supervised training samples. The evolution classification module is used to provide personalized learning intervention assistance for the objects to be identified based on the pre-trained supervised classification model.

Images